Geomatics as support to remote sensing data analysis from UAV technology using GIS open source platforms

The recent years saw a growing usage of remote sensing platforms, such as satellites and unmanned aircraft vehicles (UAVs). The Copernicus observation programme is an example of a new satellite constellation improving and promoting easy free access dataset, timely updated. The high popularity of UAVs is related to the obtaining high-resolution data, quickly at a relatively low cost. As result, sensors with high spatial and temporal resolution produce a great number of data, and these data increase exponentially. Consequently, the software for image processing play a key role in the diffusion of this technology. The satellites take advantage of dedicated software for imagery analysis. The UAVs use structure from motion techniques for photogrammetric processing. However, the data analysis for both sensors is based on the classic pixel-based and object-based remote sensing techniques. Moreover, there is a growing demand for innovative tools to analyse huge dataset and to integrate the information for environmental analyses, monitor geomorphological aspects and land use studies, in particular in rural areas. This thesis aims to study how GIS software using open source libraries can integrate information extracted from the satellite and the UAVs imagery, using a machine learning approach in a multilevel remote sensing framework. The main research questions are: (1) Can the classic techniques of remote sensing be used to extract suitable land use/land cover (LULC) maps – suitable in terms of classification accuracy – for the very high-resolution imagery of UAVs? (2) Can information from images of UAVs be merged with data from satellite images in the same area to achieve better results? (3) Which methods are optimal to analyse imagery of UAVs, and which benefits can be achieved through the use of more sophisticated techniques, such as the integration of multisource spatial information? To answer the research questions, a multi-level framework has been developed to integrate the information derived from remote sensing techniques. The framework has been implemented using R cran libraries, and it includes a machine learning benchmark as an alternative to pixel-based and object-based approach. The benchmark allows for testing several algorithms, in terms of accuracy and processing time for classifying LULC maps. The thesis presents the result of five papers, and the main findings relating to the major research questions can be summarized as follows: (1) The classic remote sensing techniques can be applied to UAVs high-resolution imagery to obtain a fast image classification. The maximum likelihood algorithm has a better result than the minimum distance algorithm in terms of accuracy.8 (2) It is possible to integrate the satellite and UAVs temporal series. The scale affects the size of the training areas. Thus, to integrate the satellite and UAV information, the size of the regions of interest (ROIs) shall be larger than the ground sample distance (GSD) of the satellite. The use of large ROIs can avoid the noise from nearby areas. In addition, to limit the noise due to high-resolution images, the value of the digital number (DN) inside the ROIs should be homogenous. (3) The machine learning can be applied to both satellite and UAVs imagery and integrate spatial information. The dataset derived from high-resolution imagery can be considered as big data paradigm, in terms of data size and the processing time. Using a subset greater than the 8% of the total is possible to have a good results (kappa score ranges between 80% and 90%) and fast processing time. In addition, sensitivity analysis can help to define the contribution of each layer of the multi-level framework